2026 Theses Doctoral
Data-Driven Methods for Estimating and Forecasting Infectious Disease Transmission Under Nonstationary Dynamics
In this dissertation, we develop and evaluate data-driven methods for estimating and forecasting infectious disease dynamics under nonstationary conditions. Two complementary problems are addressed: estimating the time-varying basic reproduction number 𝑅_𝑡 during abrupt transmission shifts and extracting leading indicators of hospitalization surges from human mobility networks.
In the first part of this dissertation, we address the challenge of tracking sudden changes in disease dynamics. Current data assimilation methods often underestimate uncertainty and struggle to capture temporal trends when there are abrupt changes in parameters. This can lead to filter divergence, where the model becomes overconfident in its prior and fails to track the true trajectory. Chapter 2 evaluates six methods for inferring 𝑅_𝑡: the Ensemble Adjustment Kalman Filter (EAKF) with no inflation, fixed inflation, and a data-driven adaptive inflation scheme; the Ensemble Square Root Smoother (EnSRS) with adaptive inflation; EpiEstim; and EpiFilter. Performance is assessed on synthetic data generated by a stochastic Susceptible-Infected-Recovered (SIR) model with sigmoidal 𝑅_𝑡 transitions and on empirical COVID-19 case data from Manhattan during the Omicron wave. Both adaptive inflation methods outperform the others in parameter estimation accuracy and posterior predictive fit, providing a robust, data-driven alternative to manually selected inflation parameters.
The second part of this dissertation addresses the challenge of predicting the onset and peak timing of epidemic outbreak waves. Operational forecast models exhibit limitations in accurately anticipating these critical points, leaving public health systems with limited lead time to respond. Spatiotemporal Graph Neural Networks (STGNNs) utilizing human mobility networks have emerged as powerful forecasting tools, but their black-box nature often obscures the specific features driving their predictions. We introduce a novel framework that pairs an STGNN with an Explainable Artificial Intelligence (XAI) algorithm to extract a sparse subgraph of mobility edges most critical for model prediction. Applying this framework to COVID-19 (Chapter 3) and seasonal influenza (Chapter 4), we demonstrate that the aggregate count of these identified edges serves as a robust leading indicator of national and regional hospitalizations. Our findings suggest this framework provides a supplemental, mobility-based early warning system, enhancing situational awareness for public health monitoring before healthcare systems become overwhelmed.
Taken together, these methodologies equip public health systems with real-time tools to navigate volatile transmission regimes. By enhancing parameter inference during abrupt changes in dynamics and extracting early warning signals from mobility data, this work helps shift disease surveillance from reactive observation to proactive mitigation.
Geographic Areas
Subjects
Files
This item is currently under embargo. It will be available starting 2027-07-14.
More About This Work
- Academic Units
- Applied Physics and Applied Mathematics
- Thesis Advisors
- Shaman, Jeffrey L.
- Spiegelman, Marc W.
- Degree
- Ph.D., Columbia University
- Published Here
- October 21, 2026
Notes
Mathematics, Epidemiology--Mathematical models, Kalman filtering, Neural networks (Computer science), COVID-19 (Disease)
Additional thesis advisor(s): Spiegelman, Marc W.